MemoryCache: Augmenting local AI with browser data

Original Article ↗ Hacker News Discussion ↗

Mozilla’s experimental MemoryCache project aims to run AI models directly on users’ devices and use locally stored browser data—like history, bookmarks, and page content—to power search, summarization, and assistant-like features without sending data to the cloud. Commenters are broadly enthusiastic about the “models come to the data” approach for privacy, personalization, and better retrieval, but raise questions about hardware requirements, archiving and search UX, ad pollution, and whether Mozilla should prioritize such AI features over core browser improvements. Overall, it’s seen as a promising direction for private, on-device AI agents, echoing older ideas like desktop search and personal knowledge bases but updated with modern LLMs.

Models Near Data vs Data to Models

  • Strong enthusiasm for the idea that models should run locally and come to user data (history, documents, media) instead of sending data to remote services.
  • Some envision personal “secretary”-like assistants and networks of private AIs that can optionally pair with each other, contrasted with centralized, ad-driven AI.

Local LLM Feasibility & Hardware

  • Clarification that most use cases don’t require training; retrieval-augmented generation (RAG) over local data is sufficient.
  • Users report decent performance for 7B‑parameter models on modest laptops; larger or multimodal models require stronger GPUs.
  • Discussion of VRAM needs and quantized models; community forums are recommended for concrete “what can I run on my hardware?” questions.

Beyond Text: Images, Video, and Archiving

  • Some note existing local multimodal models and captioning tools; others want better video/image support and full-page archiving.
  • Multiple extensions and tools are cited for DOM snapshots, WARC-based archiving, and saving pages as single HTML files.
  • Frustration that modern browsers make robust offline page capture and media download surprisingly hard.

Desired Use Cases

  • Chatting with browsing history, better local search over pages, downloads, and bookmarks.
  • Automatic summaries of frequently visited sites and “Spotify Wrapped”-style usage views.
  • Automation: logging in, filling forms, buying tickets, clicking “skip ad” or cookie banners.
  • Some only want improved bookmark/content indexing and fuzzy search, not full-blown chat.

Privacy, Ads, and Data Integrity

  • Local models are seen as a promising privacy-preserving alternative to cloud AI, but some distrust the claim and dislike having to monitor yet another data collector.
  • Concern that ingesting ads could pollute the model or enable manipulation, similar to how advertisers target humans.

Views on Mozilla & Product Direction

  • Mixed reactions: praise for privacy-preserving, on-device experiments (citing offline translation) versus criticism that Mozilla is trend-chasing and neglecting core browser issues.
  • Debate over whether such niche features help Firefox’s market share or distract from broader user needs.
  • Some find the “MemoryCache” name misleading and the project communication vague or confusing.